VLALight-4B

VLALight: A Vision-Language-Action Model for Traffic Signal Control.

Overview

VLALight is a vision-language-action model for multi-intersection traffic signal control, built on Qwen3.5-4B. It processes visual traffic observations, incorporates information from neighboring intersections, and produces traffic-signal decisions with adaptive Fast/Slow reasoning.

This repository provides a merged checkpoint for inference. Training uses supervised fine-tuning followed by cooperative reinforcement learning with local and network-level traffic feedback.

Inference Workflow

The model is used within a two-stage control pipeline:

1. Visual perception: process intersection videos and available coordination frames to extract traffic information.
2. Cooperative decision-making: combine local and neighboring-intersection information, select Fast or Slow reasoning, and output a signal decision within <signal> tags.

The supported phase labels are ETWT, NTST, ELWL, and NLSL. Their mapping to simulator actions must follow the road-network configuration provided by the project.

Usage

Download the checkpoint:

hf download anonymous-work/VLALight-4B --local-dir ./VLALight-4B

Serve the downloaded model using a compatible OpenAI-style chat-completion server. In the project code, configure the inference endpoint and use its served model ID with run_video.py --decision-model.

The evaluation pipeline additionally requires SUMO, the video-rendering environment, scenario data, and the project's prompts and phase mappings. Model weights alone do not run a traffic simulation.

Evaluation Scope and Limitations

The project evaluates traffic control on seven traffic-flow datasets from Jinan, Hangzhou, and New York using a video-enabled traffic simulator.

This checkpoint is intended for research and simulation. Performance under real-world camera conditions and deployment on physical traffic infrastructure have not been established. Outputs require validation before any safety-critical use.

License

This repository is distributed under Apache-2.0. Users must also comply with the applicable terms of the base model and any separately obtained data or simulation assets.